VOLUME 78, NUMBER 10

PHYSICAL REVIEW LETTERS

10 MARCH 1997

Heat Conduction in Chains of Nonlinear Oscillators

Stefano Lepri,1,3 Roberto Livi,1,3 and Antonio Politi2
1

Dipartimento di Fisica, Universit di Bologna and Istituto Nazionale di Fisica Nucleare, I-40127 Bologna, Italy
2
Istituto Nazionale di Ottica and Istituto Nazionale di Fisica Nucleare, I-50125 Firenze, Italy
3
Istituto Nazionale di Fisica della Materia, I-50125 Firenze, Italy
(Received 22 October 1996; revised manuscript received 30 January 1997)
We numerically study heat conduction in chains of nonlinear oscillators with time-reversible
thermostats. A nontrivial temperature profile is found to set in, which obeys a simple scaling
relation for increasing the number N of particles. The thermal conductivity diverges approximately
as N 1y2 , indicating that chaotic behavior is not enough to ensure the Fourier law. Finally, we show
that the microscopic dynamics ensures fulfillment of a macroscopic balance equation for the entropy
production. [S0031-9007(97)02611-2]
PACS numbers: 44.10.+i, 05.45.+b, 05.60.+w, 05.70.Ln

The approach to nonequilibrium statistical mechanics

through the introduction of microscopically timereversible models has been shown to be rather powerful
in the context of many-particle dynamics [1]. If the
reversibility property is supplemented by the so-called
chaotic hypothesis, the tools developed for strictly
hyperbolic systems allowed making general statistical
predictions that have been successfully tested [2]. Among
the achievements of this approach, we recall the derivation
of the Onsager reciprocity relations [3] and the expression of entropy production in terms of a self-generated
dissipation rate [4].
So far, however, most of the numerical efforts in
this area have been restricted mainly to the description
of gases and fluids, where the thermostats, introduced
to keep the energy constant, affect each particle [1]
(with a few exceptions, such as Refs. [3,5,6]). In the
present Letter we investigate the possibility of extending
the above approach to a chain of coupled nonlinear
oscillators, with specific reference to heat-conduction
properties of insulating solids. In this context, the most
natural choice is to put only the chain extrema in contact
with two thermal baths at different temperatures.
A further motivation for the present work is the lack
of convincing results about the validity of the Fourier
conduction law in 1D systems. Let us briefly review
the current state of the arts. In the simplest case of
coupled harmonic oscillators, it was rigorously shown
[7] that, if the extrema of the chain are put in contact
with stochastic heat reservoirs operating at different
temperatures, a nonequilibrium stationary state sets in
with no temperature gradient in the bulk. As a result,
thermal conductivity k turns out to be proportional to
the number of oscillators N. Such a divergence simply
follows from the existence of extended waves (phonons)
freely traveling and carrying energy along the lattice
without attenuation.
Afterwards, the role of impurities has been taken into
account, since it was expected that phonon waves should

be damped by the scattering processes due to defects, thus

possibly removing the divergence of k. Unfortunately, it
was found that although isotopic disorder in a harmonic
chain yields a nonzero temperature gradient in the bulk
[8,9], it still implies a diverging conductivity sk N 1y2 d
[10,11]. A finite k has been obtained only by placing all
the oscillators in contact with independent thermal baths
[12]. As a result, one can conclude that no physically
sound description of Fourier law can be obtained with
harmonic chains.
More than disorder, anharmonicity has been invoked
as the key feature of real solids responsible for normal
heat conduction [13]: Nonlinearities make phonons interact among themselves, thus impeding free propagation.
In the spirit of the general theory of dynamical systems,
nonintegrability, rather than anharmonicity, is the property that should be responsible for a finite conductivity.
In fact, nonlinear normal modes (solitons) freely transport
energy along the chain. Our numerical simulations performed with a Toda chain [see Eqs. (1) and (2) for the
precise definition of the model] do reveal the same scenario as for linear chains (see also [14]).
Numerical experiments for chains with chaotic smooth
potentials [14,15] have been performed with too few
particles to allow, even in the most detailed investigation
[16], a conclusive study of the dependence of k on
N. The same can be said for the case where both
anharmonicity and disorder have been simultaneously
included [17,18].
Finally, we must recall two somehow artificial models
that lead to contradictory conclusions: The first one is a
chain of harmonic oscillators with an infinite barrier set at
a given distance [19]; the second is the so-called ding-aling model, where harmonic oscillators alternate with free
particles [20]. While in the former case the conductivity
has again been found to diverge, in the latter the authors
found convincing evidence that it attains a finite value.
In this Letter we study the Fermi-Pasta-Ulam (FPU)
model, which represents the simplest anharmonic

1896

1997 The American Physical Society

0031-9007y97y78(10)y1896(4)$10.00

VOLUME 78, NUMBER 10

PHYSICAL REVIEW LETTERS

10 MARCH 1997

approximation of a monoatomic solid. Specifically, we

consider a chain of N oscillators, indicating with qi
the displacement of the ith particle from its equilibrium position. Fixed boundary conditions are assumed
sq0 qN11 0d, while the dynamics of the central
N 2 2 oscillators is ruled by the equations of motion
q i fi 2 fi11 ,
2V 0 sqi

(1)
x 2 y2

2 qi21 d and V sxd

1
where fi
is the interaction potential ( b has been fixed equal to
0.1). Nos-Hoover thermostats [1,21] act on the first and
the last particle, keeping them at temperature T1 and T2 ,
respectively,
q 2
q 1 2z1 q 1 1 f1 2 f2 , z1 1 2 1 ,
T1
q N 2z2 q N 1 fN 2 fN11 ,
q 2
(2)
z2 N 2 1 .
T2

bx 4 y4

The dynamical equations are left invariant under time reversal composed with the involution pi ! 2pi . Recent
numerical observations [22] show that time-reversible
nonequilibrium dynamics yields results compatible with
the predictions of Ref. [2], despite the fact that the
system under investigation is not strictly Anosov. We
expect that this should hold also for our model at sufficiently high temperatures. However, we shall not further
address this point here; this will be the subject of a forthcoming paper [23].
We have performed extensive numerical simulations
with several values of N and T6 , integrating the equations
of motion with an improved fourth-order Runge-KuttaGhil algorithm. The first clear result is the convergence
to a well defined spatial profile of the local temperature
Ti k pi2 l (k?l denoting time average). The asymptotic
stationary state satisfies the local equilibrium condition, as
confirmed by the fluctuations of Ti that are in agreement
with the canonical ones. The only exceptions are represented by the particles close to the boundaries, where the
temperature profile seems to exhibit a singularity. Globally, the profiles satisfy a simple scaling relation, as clearly
shown in Fig. 1, where the values of Ti , corresponding to
different chain lengths (and the same boundary temperatures), are plotted versus iyN. The adoption of the above
scaled units is tantamount to considering the continuum limit with the lattice spacing a equal to 1yN. However, this is to be taken only as a formal interpretation, as
the mass density obviously diverges when N ! `; conversely, if the equations are rescaled in such a way that
both energy and mass densities are kept constant, one finds
that the nonlinearity coefficient b should diverge.
The nonlinear shape of the profiles could be interpreted
as an indication of a temperature-dependent conductivity,
but this is incorrect, since simulations done with such small
temperature differences as T1 2 T2 4 still reveal clear
deviations from linearity. This is rather an indication of

FIG. 1. Scaling of the temperature profiles for the FPU b

model. The imposed temperatures are T1 152 and T2 24,
and chain lengths are N 128, 194, and 256 (dashed, dotted,
and solid lines, respectively). Averages are carried over a time
interval 106 , after a transient 104 .

the relevant role played by boundary conditions; indeed, a

seemingly square-root-type singularity in the temperature
profile is always observed at the chain extrema.
The next result concerns the local heat flux Jsx, td,
which is implicitly defined by the continuity equation,

Hsx,
td 1 divJsx, td 0 ,
(3)
P
where H i Hi dsx 2 xi d, Hi pi2 y2 1 V sqi 2
qi21 d and xi ia 1 qi . By Fourier transforming (in
space) Eq. (3), and upon expanding in powers of the
wave number k, one eventually finds that the heat flux at
the ith position is given by [23,24]
Ji std api fi11 ,

(4)

where pi fi11 has the simple interpretation of the flow of

potential energy from the ith to the neighboring particle.
We have checked that J ; kJi stdl is independent of the
lattice position i, as it should indeed be for a stationary
nonequilibrium state.
The only physically meaningful setting for the comparison of heat fluxes for different values of N is achieved by
fixing a 1, as it is the case in real systems where the
lattice spacing is determined by the mutual interactions.
The data reported in Fig. 2 shows that J scales to zero as
N 2a , with a 0.55 6 0.05. The same scaling behavior
has been obtained for different choices of the temperatures
T1 and T2 , provided that they are sufficiently large
to ensure a chaotic behavior. This implies that the
conductivity,
k

J
,
dT ydx

(5)

diverges as N 12a , since the temperature gradient vanishes

as N 21 . Therefore, we are forced to conclude that
1897

VOLUME 78, NUMBER 10

PHYSICAL REVIEW LETTERS

10 MARCH 1997

Returning to the usual case, the energy balance at the

chain extrema implies that
2
J 2kz6 p1,N
l 2kz6 lT6 ,

FIG. 2. Scaling of the heat flux J with the number of

oscillators N for the FPU b model (same temperatures as in
Fig. 1). The inset refers to the case of an imposed constant
gradient sT1 2 T2 dyN (see text) with the same boundary
temperatures: Scaling with N 21 implies that the conductivity
is constant.

Fourier law is not satisfied in the present framework

and that chaoticity is not sufficient to ensure its validity.
Surprisingly, the above behavior is similar to the one
found in harmonic chains with random masses [11],
as if disorder and anharmonicity played the same role.
However, we have no explanation for this fact.
Two further remarks should be added as a comment to
the scaling behavior of J. First, from the very definitions
[Eqs. (4) and (5)], one realizes that the assumption a
N 21y4 implies an asymptotically finite k, but this is no
more than just a formal statement. Second, notice that
in the present philosophy, which is the standard one
adopted in the literature, T1 and T2 are kept fixed while
N diverges, so that the temperature gradient (i.e., the
external field) goes to 0. Accordingly, in the limit of
large N, the chain gets closer and closer to equilibrium
so that, independently of sT1 2 T2 d, a linear regime
(in the Green-Kubo sense) is eventually attained. This
is at variance with other physical settings, such as
electric charge transport, where the external field is a free
parameter whose magnitude can be fixed independently
of the system size. Accordingly, it is not obvious how to
study nonlinear corrections in this framework, if they are
relevant at all.
As a last comment on thermal conductivity in FPU
chains, we want to stress that a truly finite k is observed
when each particle is thermostated independently, according to a linear temperature profile. Obviously, in this case,
kJi stdl depends on the lattice site, but its average value
over all sites is found to scale as N 21 (see the inset of
Fig. 2). This is analogous to what was found in Ref. [12]
with stochastic heat baths. Needless to say, this result
sounds a bit artificial, as the profile is imposed from the
outside.
1898

(6)

where the last equality is obtained from the condition

kdz62 ydtl 0. The above equation expresses the general
scenario arising in time-reversible models that a nonequilibrium stationary state corresponds to a spontaneous
emergence of dissipation [2,4]. The global volume contraction rate g in phase space is given by the average of
the divergence of the velocity field, i.e., g kz1 1 z2 l.
In all simulations we checked that g . 0, as long as
T1 fi T2 , consistently with a theorem recently proved by
Ruelle [25]. In any case, kz1 l is always negative (provided that T1 . T2 ), as indeed prescribed by energy balance. In fact, the energy is pumped in from the hot reservoir, flows through the chain, and is eventually absorbed
in the cold reservoir. Dynamically, it is at least bizarre
that the hot thermostat is characterized by a local expansion of volumes: This is completely opposite to the approach in terms of stochastic baths, where dissipation is
always assumed. To what extent this peculiar feature is
physically meaningful is unclear; nevertheless, the interpretation in terms of entropy production makes perfect
sense. In fact, Eq. (6) can be rewritten as

1
1
,
(7)
2
kz1 l 1 kz2 l J
T2
T1
with the convention that J . 0 is an incoming flux.
Equation (7) can be physically interpreted as a balance
relation for the global entropy production. According to
the general principles of irreversible thermodynamics, the
local rate of entropy production s in the bulk is given by

d
1
ssxd J
.
(8)
dx Tsxd
Upon integrating Eq. (8), the right-hand side of Eq. (7)
is obtained, which can thus be interpreted as the global
production rate of entropy in the bulk. On the other hand,
according to general arguments on reversible thermostats
[2], the left-hand side of Eq. (7) is identified with the
entropy production from the heat baths. Equation (7) has
been numerically tested in a wide range of temperatures.
A relevant consequence of Eq. (6) is that z6 are
proportional to J, so that not only the fluxes but also
the dissipation g vanish in the thermodynamic limit
N ! `. This is indeed a remarkable difference with
respect to other models of gases and fluids studied, e.g.,
in Refs. [2,4,5], where the dissipation is always extensive.
In our opinion, this is due to the vanishing of dT ydx, and
not to the fact that thermostats act only at the boundaries.
Indeed, by globally thermostating the two halves of the
lattice at two different temperatures [26], we find that g,

VOLUME 78, NUMBER 10

PHYSICAL REVIEW LETTERS

which is now the sum of N contributions, still goes to zero

as N 21y2 .
Although it is generally believed that nonlinearity
yields a finite thermal transport coefficient, we have
shown that this is not true in FPU chains equipped
with time-reversible thermal baths. The specific choice
of the thermostat does not affect the generality of our
conclusions since the repetition of some simulations
with isokinetic (Gaussian) thermostats [1] lead to similar
results. Another possible explanation of the divergence
of k could be the dimensionality of the system: It might
be that solitonlike propagation is generically favored
in 1D systems. At the present stage this is only a
speculation that needs further investigation. Another item
that should be seriously taken into account concerns
boundary conditions. If, as we suspect, the nonlinear
shape of the profile stems from the peculiar functioning
of the thermostats, it is not to be excluded that a different
setup could lead to more physically meaningful results.
Some preliminary simulations performed by thermostating
a number of O sNd particles on both the left and right
sides suggest a possible slow convergence to a finite
k value. In any case, the anomalous behavior of the
conductivity is presumably to be attributed to the almost
free evolution of the small-wave-number Fourier modes
even at high energies [27]. In fact, the small coupling of
such modes with the rest of the chain suggests the possible
existence of perturbations traveling almost undamped
along the chain.
We want to thank G. Gallavotti for encouraging us
to undertake the present study and for enlightening
discussions, and E. G. D. Cohen for his many suggestions,
criticisms, and transfer of enthusiasm. One of us (R. L.) is
indebted to Ph. Choquard for clarifying discussions on the
definition of heat flux. Part of this work has been done at
the Institute for Scientific Interchange, Turin, Italy, during
the Workshop of the EEC Network on Complexity and
Chaos.

[1] D. J. Evans and G. P. Morriss Statistical Mechanics of

Nonequilibrium Liquids (Academic Press, San Diego,
1990).

10 MARCH 1997

[2] G. Gallavotti and E. G. D. Cohen, J. Stat. Phys. 80, 931

(1995); Phys. Rev. Lett. 74, 2694 (1995).
[3] G. Gallavotti, J. Stat. Phys. (to be published).
[4] D. J. Evans, E. G. D. Cohen, and G. P. Morriss, Phys. Rev.
Lett. 71, 2401 (1993).
[5] H. A. Posch and W. G. Hoover, Phys. Rev. A 39, 2175
(1989).
[6] N. I. Chernov and J. L. Lebowitz (to be published).
[7] Z. Rieder, J. L. Lebowitz, and E. Lieb, J. Math. Phys. 8,
1073 (1967).
[8] A. Casher and J. L. Lebowitz, J. Math. Phys. 12, 1701
(1971).
[9] A. J. OConnor and J. L. Lebowitz, J. Math. Phys. 15, 692
(1974).
[10] H. Matsuda and K. Ishii, Prog. Theor. Phys. Suppl. 45, 56
(1970).
[11] See J. B. Keller, G. C. Papanicolaou, and J. Weilenmann,
Commun. Pure Appl. Math. 32, 583 (1978), where open
boundary conditions are assumed. However, one should
stress that different choices, such as fixed conditions,
appear to yield a different behavior (see also [19]).
[12] M. Bolsterli, M. Rich, and W. M. Visscher, Phys. Rev. A
1, 1086 (1970).
[13] R. E. Peierls, Quantum Theory of Solids (Oxford University Press, London, 1955).
[14] F. Mokross and H. Bttner, J. Phys. C 16, 4539 (1983).
[15] N. Nakazawa, Prog. Theor. Phys. Suppl. 45, 231 (1970).
[16] H. Kaburaki and M. Machida, Phys. Lett. A 181, 85
(1993).
[17] D. N. Payton, M. Rich, and W. M. Visscher, Phys. Rev.
160, 706 (1967).
[18] E. A. Jackson, J. R. Pasta, and J. F. Waters, J. Comput.
Phys. 2, 207 (1968).
[19] W. M. Visscher, Methods in Computational Physics (Academic Press, New York, 1976), Vol. 15, p. 371.
[20] G. Casati, J. Ford, F. Vivaldi, and W. M. Visscher, Phys.
Rev. Lett. 52, 1861 (1984).
[21] S. Nos, J. Chem. Phys. 81, 511 (1984); W. G. Hoover,
Phys. Rev. A 31, 1695 (1985).
[22] F. Bonetto, G. Gallavotti, and P. Garrido (to be published).
[23] S. Lepri, R. Livi, and A. Politi (unpublished).
[24] Ph. Choquard, Helv. Phys. Acta 36, 415 (1963).
[25] D. Ruelle (to be published).
[26] H. A. Posch, W. Hoover, and L. W. Campbell, Chaos 3,
325 (1993).
[27] C. Alabiso, M. Casartelli, and P. Marenzoni, J. Stat. Phys.
79, 451 (1995).

1899